fixes for fp context switching

[unix-history] / usr / src / sys / kern / kern_clock.c

diff --git a/usr/src/sys/kern/kern_clock.c b/usr/src/sys/kern/kern_clock.c
index 3ccf1df..d979a48 100644 (file)
--- a/usr/src/sys/kern/kern_clock.c
+++ b/usr/src/sys/kern/kern_clock.c
@@ -1,63 +1,77 @@
-/*     kern_clock.c    4.42    82/10/21        */
+/*-
+ * Copyright (c) 1982, 1986, 1991 The Regents of the University of California.
+ * All rights reserved.
+ *
+ * %sccs.include.redist.c%
+ *
+ *     @(#)kern_clock.c        7.16 (Berkeley) %G%
+ */
+
+#include "param.h"
+#include "systm.h"
+#include "dkstat.h"
+#include "callout.h"
+#include "kernel.h"
+#include "proc.h"
+#include "resourcevar.h"

 
 

-#include "../h/param.h"
-#include "../h/systm.h"
-#include "../h/dk.h"
-#include "../h/callout.h"
-#include "../h/dir.h"
-#include "../h/user.h"
-#include "../h/kernel.h"
-#include "../h/proc.h"
-#include "../h/psl.h"
-#include "../h/vm.h"
-#include "../h/text.h"
-#ifdef MUSH
-#include "../h/quota.h"
-#include "../h/share.h"
+#include "machine/cpu.h"
+
+#ifdef GPROF
+#include "gprof.h"

 #endif
 
 #endif
 

+#define ADJTIME                /* For now... */
+#define        ADJ_TICK 1000
+int    adjtimedelta;
+

 /*
  * Clock handling routines.
  *
 /*
  * Clock handling routines.
  *

- * This code is written for a machine with only one interval timer,
- * and does timing and resource utilization estimation statistically
- * based on the state of the machine hz times a second.  A machine
- * with proper clocks (running separately in user state, system state,
- * interrupt state and idle state) as well as a time-of-day clock
- * would allow a non-approximate implementation.
+ * This code is written to operate with two timers which run
+ * independently of each other. The main clock, running at hz
+ * times per second, is used to do scheduling and timeout calculations.
+ * The second timer does resource utilization estimation statistically
+ * based on the state of the machine phz times a second. Both functions
+ * can be performed by a single clock (ie hz == phz), however the 
+ * statistics will be much more prone to errors. Ideally a machine
+ * would have separate clocks measuring time spent in user state, system
+ * state, interrupt state, and idle state. These clocks would allow a non-
+ * approximate measure of resource utilization.

  */
 
 /*
  * TODO:
  */
 
 /*
  * TODO:

- *     * Keep more accurate statistics by simulating good interval timers.
- *     * Use the time-of-day clock on the VAX to keep more accurate time
- *       than is possible by repeated use of the interval timer.
- *     * Allocate more timeout table slots when table overflows.
+ *     time of day, system/user timing, timeouts, profiling on separate timers
+ *     allocate more timeout table slots when table overflows.

  */
 
  */
 

-/* bump a timeval by a small number of usec's */
-#define        bumptime(tp, usec) \
-       (tp)->tv_usec += usec; \
-       if ((tp)->tv_usec >= 1000000) { \
-               (tp)->tv_usec -= 1000000; \
-               (tp)->tv_sec++; \
-       }
+/*
+ * Bump a timeval by a small number of usec's.
+ */
+#define BUMPTIME(t, usec) { \
+       register struct timeval *tp = (t); \
+ \
+       tp->tv_usec += (usec); \
+       if (tp->tv_usec >= 1000000) { \
+               tp->tv_usec -= 1000000; \
+               tp->tv_sec++; \
+       } \
+}

 
 /*
 
 /*

- * The (single) hardware interval timer.
- * We update the events relating to real time, and then
- * make a gross assumption: that the system has been in the
- * state it is in (user state, kernel state, interrupt state,
- * or idle state) for the entire last time interval, and
- * update statistics accordingly.
+ * The hz hardware interval timer.
+ * We update the events relating to real time.
+ * If this timer is also being used to gather statistics,
+ * we run through the statistics gathering routine as well.

  */
  */

-/*ARGSUSED*/
-hardclock(pc, ps)
-       caddr_t pc;
+hardclock(frame)
+       clockframe frame;

 {
        register struct callout *p1;
 {
        register struct callout *p1;

-       register struct proc *p;
-       register int s, cpstate;
+       register struct proc *p = curproc;
+       register struct pstats *pstats;
+       register int s;

 
        /*
         * Update real-time timeout queue.
 
        /*
         * Update real-time timeout queue.


@@ -69,10 +83,43 @@ hardclock(pc, ps)
         * Decrementing just the first of these serves to decrement the time
         * to all events.
         */
         * Decrementing just the first of these serves to decrement the time
         * to all events.
         */

-       for (p1 = calltodo.c_next; p1 && p1->c_time <= 0; p1 = p1->c_next)
-               --p1->c_time;
-       if (p1)
-               --p1->c_time;
+       p1 = calltodo.c_next;
+       while (p1) {
+               if (--p1->c_time > 0)
+                       break;
+               if (p1->c_time == 0)
+                       break;
+               p1 = p1->c_next;
+       }
+
+       /*
+        * Curproc (now in p) is null if no process is running.
+        * We assume that curproc is set in user mode!
+        */
+       if (p)
+               pstats = p->p_stats;
+       /*
+        * Charge the time out based on the mode the cpu is in.
+        * Here again we fudge for the lack of proper interval timers
+        * assuming that the current state has been around at least
+        * one tick.
+        */
+               /*
+                * CPU was in user state.  Increment
+                * user time counter, and process process-virtual time
+                * interval timer. 
+                */
+               BUMPTIME(&p->p_utime, tick);
+               if (timerisset(&pstats->p_timer[ITIMER_VIRTUAL].it_value) &&
+                   itimerdecr(&pstats->p_timer[ITIMER_VIRTUAL], tick) == 0)
+                       psignal(p, SIGVTALRM);
+       } else {
+               /*
+                * CPU was in system state.
+                */
+               if (p)
+                       BUMPTIME(&p->p_stime, tick);
+       }

 
        /*
         * If the cpu is currently scheduled to a process, then
 
        /*
         * If the cpu is currently scheduled to a process, then


@@ -82,46 +129,109 @@ hardclock(pc, ps)
         * This assumes that the current process has been running
         * the entire last tick.
         */
         * This assumes that the current process has been running
         * the entire last tick.
         */

-       if (!noproc) {
-               s = u.u_procp->p_rssize;
-               u.u_ru.ru_idrss += s; u.u_ru.ru_isrss += 0;     /* XXX */
-               if (u.u_procp->p_textp) {
-                       register int xrss = u.u_procp->p_textp->x_rssize;
+       if (p) {
+               if ((p->p_utime.tv_sec+p->p_stime.tv_sec+1) >
+                   p->p_rlimit[RLIMIT_CPU].rlim_cur) {
+                       psignal(p, SIGXCPU);
+                       if (p->p_rlimit[RLIMIT_CPU].rlim_cur <
+                           p->p_rlimit[RLIMIT_CPU].rlim_max)
+                               p->p_rlimit[RLIMIT_CPU].rlim_cur += 5;
+               }
+               if (timerisset(&pstats->p_timer[ITIMER_PROF].it_value) &&
+                   itimerdecr(&pstats->p_timer[ITIMER_PROF], tick) == 0)
+                       psignal(p, SIGPROF);
+
+               /*
+                * We adjust the priority of the current process.
+                * The priority of a process gets worse as it accumulates
+                * CPU time.  The cpu usage estimator (p_cpu) is increased here
+                * and the formula for computing priorities (in kern_synch.c)
+                * will compute a different value each time the p_cpu increases
+                * by 4.  The cpu usage estimator ramps up quite quickly when
+                * the process is running (linearly), and decays away
+                * exponentially, * at a rate which is proportionally slower
+                * when the system is busy.  The basic principal is that the
+                * system will 90% forget that a process used a lot of CPU
+                * time in 5*loadav seconds.  This causes the system to favor
+                * processes which haven't run much recently, and to
+                * round-robin among other processes.
+                */
+               p->p_cpticks++;
+               if (++p->p_cpu == 0)
+                       p->p_cpu--;
+               if ((p->p_cpu&3) == 0) {
+                       setpri(p);
+                       if (p->p_pri >= PUSER)
+                               p->p_pri = p->p_usrpri;
+               }
+       }
+
+       /*
+        * If the alternate clock has not made itself known then
+        * we must gather the statistics.
+        */
+       if (phz == 0)
+               gatherstats(&frame);

 
 

-                       s += xrss;
-                       u.u_ru.ru_ixrss += xrss;
+       /*
+        * Increment the time-of-day, and schedule
+        * processing of the callouts at a very low cpu priority,
+        * so we don't keep the relatively high clock interrupt
+        * priority any longer than necessary.
+        */
+#ifdef ADJTIME
+       if (adjtimedelta == 0)
+               bumptime(&time, tick);
+       else {
+               if (adjtimedelta < 0) {
+                       bumptime(&time, tick-ADJ_TICK);
+                       adjtimedelta++;
+               } else {
+                       bumptime(&time, tick+ADJ_TICK);
+                       adjtimedelta--;

                }
                }

-               if (s > u.u_ru.ru_maxrss)
-                       u.u_ru.ru_maxrss = s;
-               if ((u.u_ru.ru_utime.tv_sec+u.u_ru.ru_stime.tv_sec+1) >
-                   u.u_rlimit[RLIMIT_CPU].rlim_cur) {
-                       psignal(u.u_procp, SIGXCPU);
-                       if (u.u_rlimit[RLIMIT_CPU].rlim_cur <
-                           u.u_rlimit[RLIMIT_CPU].rlim_max)
-                               u.u_rlimit[RLIMIT_CPU].rlim_cur += 5;
+       }
+#else
+       if (timedelta == 0)
+               BUMPTIME(&time, tick)
+       else {
+               register delta;
+
+               if (timedelta < 0) {
+                       delta = tick - tickdelta;
+                       timedelta += tickdelta;
+               } else {
+                       delta = tick + tickdelta;
+                       timedelta -= tickdelta;

                }
                }

-               if (timerisset(&u.u_timer[ITIMER_PROF].it_value) &&
-                   itimerdecr(&u.u_timer[ITIMER_PROF], tick) == 0)
-                       psignal(u.u_procp, SIGPROF);
+               BUMPTIME(&time, delta);

        }
        }

+#endif
+       setsoftclock();
+}
+
+int    dk_ndrive = DK_NDRIVE;
+/*
+ * Gather statistics on resource utilization.
+ *
+ * We make a gross assumption: that the system has been in the
+ * state it is in (user state, kernel state, interrupt state,
+ * or idle state) for the entire last time interval, and
+ * update statistics accordingly.
+ */
+gatherstats(framep)
+       clockframe *framep;
+{
+       register int cpstate, s;

 
        /*
 
        /*

-        * Charge the time out based on the mode the cpu is in.
-        * Here again we fudge for the lack of proper interval timers
-        * assuming that the current state has been around at least
-        * one tick.
+        * Determine what state the cpu is in.

         */
         */

-       if (USERMODE(ps)) {
+       if (CLKF_USERMODE(framep)) {

                /*
                /*

-                * CPU was in user state.  Increment
-                * user time counter, and process process-virtual time
-                * interval timer.
+                * CPU was in user state.

                 */
                 */

-               bumptime(&u.u_ru.ru_utime, tick);
-               if (timerisset(&u.u_timer[ITIMER_VIRTUAL].it_value) &&
-                   itimerdecr(&u.u_timer[ITIMER_VIRTUAL], tick) == 0)
-                       psignal(u.u_procp, SIGVTALRM);
-               if (u.u_procp->p_nice > NZERO)
+               if (curproc->p_nice > NZERO)

                        cpstate = CP_NICE;
                else
                        cpstate = CP_USER;
                        cpstate = CP_NICE;
                else
                        cpstate = CP_USER;


@@ -137,20 +247,15 @@ hardclock(pc, ps)
                 * This is approximate, but the lack of true interval
                 * timers makes doing anything else difficult.
                 */
                 * This is approximate, but the lack of true interval
                 * timers makes doing anything else difficult.
                 */

+               cpstate = CP_SYS;
+               if (curproc == NULL && CLKF_BASEPRI(framep))
+                       cpstate = CP_IDLE;

 #ifdef GPROF
 #ifdef GPROF

-               int k = pc - s_lowpc;
-               if (profiling < 2 && k < s_textsize)
-                       kcount[k / sizeof (*kcount)]++;
+               s = CLKF_PC(framep) - s_lowpc;
+               if (profiling < 2 && s < s_textsize)
+                       kcount[s / (HISTFRACTION * sizeof (*kcount))]++;

 #endif
 #endif

-               cpstate = CP_SYS;
-               if (noproc) {
-                       if ((ps&PSL_IPL) != 0)
-                               cpstate = CP_IDLE;
-               } else {
-                       bumptime(&u.u_ru.ru_stime, tick);
-               }

        }
        }

-

        /*
         * We maintain statistics shown by user-level statistics
         * programs:  the amount of time in each cpu state, and
        /*
         * We maintain statistics shown by user-level statistics
         * programs:  the amount of time in each cpu state, and


@@ -160,46 +265,6 @@ hardclock(pc, ps)
        for (s = 0; s < DK_NDRIVE; s++)
                if (dk_busy&(1<<s))
                        dk_time[s]++;
        for (s = 0; s < DK_NDRIVE; s++)
                if (dk_busy&(1<<s))
                        dk_time[s]++;

-
-       /*
-        * We adjust the priority of the current process.
-        * The priority of a process gets worse as it accumulates
-        * CPU time.  The cpu usage estimator (p_cpu) is increased here
-        * and the formula for computing priorities (in kern_synch.c)
-        * will compute a different value each time the p_cpu increases
-        * by 4.  The cpu usage estimator ramps up quite quickly when
-        * the process is running (linearly), and decays away exponentially,
-        * at a rate which is proportionally slower when the system is
-        * busy.  The basic principal is that the system will 90% forget
-        * that a process used a lot of CPU time in 5*loadav seconds.
-        * This causes the system to favor processes which haven't run
-        * much recently, and to round-robin among other processes.
-        */
-       if (!noproc) {
-               p = u.u_procp;
-               p->p_cpticks++;
-               if (++p->p_cpu == 0)
-                       p->p_cpu--;
-#ifdef MUSH
-               p->p_quota->q_cost += (p->p_nice > NZERO ?
-                   (shconsts.sc_tic * ((2*NZERO)-p->p_nice)) / NZERO :
-                   shconsts.sc_tic) * (((int)avenrun[0]+2)/3);
-#endif
-               if ((p->p_cpu&3) == 0) {
-                       (void) setpri(p);
-                       if (p->p_pri >= PUSER)
-                               p->p_pri = p->p_usrpri;
-               }
-       }
-
-       /*
-        * Increment the time-of-day, and schedule
-        * processing of the callouts at a very low cpu priority,
-        * so we don't keep the relatively high clock interrupt
-        * priority any longer than necessary.
-        */
-       bumptime(&time, tick);
-       setsoftclock();

 }
 
 /*
 }
 
 /*


@@ -207,8 +272,8 @@ hardclock(pc, ps)
  * Run periodic events from timeout queue.
  */
 /*ARGSUSED*/
  * Run periodic events from timeout queue.
  */
 /*ARGSUSED*/

-softclock(pc, ps)
-       caddr_t pc;
+softclock(frame)
+       clockframe frame;

 {
 
        for (;;) {
 {
 
        for (;;) {


@@ -217,7 +282,7 @@ softclock(pc, ps)
                register int (*func)();
                register int a, s;
 
                register int (*func)();
                register int a, s;
 

-               s = spl7();
+               s = splhigh();

                if ((p1 = calltodo.c_next) == 0 || p1->c_time > 0) {
                        splx(s);
                        break;
                if ((p1 = calltodo.c_next) == 0 || p1->c_time > 0) {
                        splx(s);
                        break;


@@ -226,33 +291,54 @@ softclock(pc, ps)
                calltodo.c_next = p1->c_next;
                p1->c_next = callfree;
                callfree = p1;
                calltodo.c_next = p1->c_next;
                p1->c_next = callfree;
                callfree = p1;

-               (void) splx(s);
+               splx(s);

                (*func)(arg, a);
        }
                (*func)(arg, a);
        }

+       /*
+        * If trapped user-mode and profiling, give it
+        * a profiling tick.
+        */
+       if (CLKF_USERMODE(&frame)) {
+               register struct proc *p = curproc;
+
+               if (p->p_stats->p_prof.pr_scale)
+                       profile_tick(p, &frame);
+               /*
+                * Check to see if process has accumulated
+                * more than 10 minutes of user time.  If so
+                * reduce priority to give others a chance.
+                */
+               if (p->p_ucred->cr_uid && p->p_nice == NZERO &&
+                   p->p_utime.tv_sec > 10 * 60) {
+                       p->p_nice = NZERO + 4;
+                       setpri(p);
+                       p->p_pri = p->p_usrpri;
+               }
+       }

 }
 
 /*
 }
 
 /*

- * Arrange that (*fun)(arg) is called in tim/hz seconds.
+ * Arrange that (*func)(arg) is called in t/hz seconds.

  */
  */

-timeout(fun, arg, tim)
-       int (*fun)();
+timeout(func, arg, t)
+       int (*func)();

        caddr_t arg;
        caddr_t arg;

-       int tim;
+       register int t;

 {
        register struct callout *p1, *p2, *pnew;
 {
        register struct callout *p1, *p2, *pnew;

-       register int t;
-       int s;
+       register int s = splhigh();

 
 

-       t = tim;
-       s = spl7();
+       if (t <= 0)
+               t = 1;

        pnew = callfree;
        if (pnew == NULL)
                panic("timeout table overflow");
        callfree = pnew->c_next;
        pnew->c_arg = arg;
        pnew = callfree;
        if (pnew == NULL)
                panic("timeout table overflow");
        callfree = pnew->c_next;
        pnew->c_arg = arg;

-       pnew->c_func = fun;
+       pnew->c_func = func;

        for (p1 = &calltodo; (p2 = p1->c_next) && p2->c_time < t; p1 = p2)
        for (p1 = &calltodo; (p2 = p1->c_next) && p2->c_time < t; p1 = p2)

-               t -= p2->c_time;
+               if (p2->c_time > 0)
+                       t -= p2->c_time;

        p1->c_next = pnew;
        pnew->c_next = p2;
        pnew->c_time = t;
        p1->c_next = pnew;
        pnew->c_next = p2;
        pnew->c_time = t;


@@ -265,16 +351,16 @@ timeout(fun, arg, tim)
  * untimeout is called to remove a function timeout call
  * from the callout structure.
  */
  * untimeout is called to remove a function timeout call
  * from the callout structure.
  */

-untimeout(fun, arg)
-       int (*fun)();
+untimeout(func, arg)
+       int (*func)();

        caddr_t arg;
 {
        register struct callout *p1, *p2;
        register int s;
 
        caddr_t arg;
 {
        register struct callout *p1, *p2;
        register int s;
 

-       s = spl7();
+       s = splhigh();

        for (p1 = &calltodo; (p2 = p1->c_next) != 0; p1 = p2) {
        for (p1 = &calltodo; (p2 = p1->c_next) != 0; p1 = p2) {

-               if (p2->c_func == fun && p2->c_arg == arg) {
+               if (p2->c_func == func && p2->c_arg == arg) {

                        if (p2->c_next && p2->c_time > 0)
                                p2->c_next->c_time += p2->c_time;
                        p1->c_next = p2->c_next;
                        if (p2->c_next && p2->c_time > 0)
                                p2->c_next->c_time += p2->c_time;
                        p1->c_next = p2->c_next;


@@ -296,7 +382,7 @@ hzto(tv)
 {
        register long ticks;
        register long sec;
 {
        register long ticks;
        register long sec;

-       int s = spl7();
+       int s = splhigh();

 
        /*
         * If number of milliseconds will fit in 32 bit arithmetic,
 
        /*
         * If number of milliseconds will fit in 32 bit arithmetic,